Predicting Spatial Distribution of Radio Signals using XGBoost

Abstract

Radio propagation models can predict the spatial distribution and strength of radio signals by simulating their propagation characteristics over coverage areas. Since the empirical propagation model with fixed structure is not suitable for complex environments due to its low prediction precision, and the ray tracing propagation model brings high cost for geographic scenario modeling, this paper proposes a data-driven radio propagation model based on machine learning. The model's input features are extracted following Non-Line-of-Sight propagation of radio signals, the scenario-spanning model structure is designed using XGBoost, and the model is trained with driving test data. We used practical measurement data collected in urban areas to evaluate the model, and it is demonstrated that the root mean square error of model prediction is no more than 10.33dB. The prediction accuracy of the proposed propagation model is better than that of empirical ones. Moreover, its prediction performance is close to that of ray tracing models, while its modeling cost is lower than that of ray tracing ones. Therefore, this model is a feasible and efficient approach for radio prediction in complex urban environment.